During the implementation of object detection frameworks, often experiments with multiple datasets are required. This involves accessing bounding box groundtruth annotations of different object categories in different datasets. This is a simple-looking task that is made a little challenging by varying formats in which different datasets organize their groundtruth annotations.
This project provides a framework to quickly read bounding box annotations of different datasets without having to worry about the actual details of native storage formats. This project has been initialized with a codebase using which a number of major pedestrian detection datasets can be easily read. It will be regularly updated to include other pedestrian and non-pedestrian datasets as well.
- Python 2.7 (with PIL package)
For a user there are just two pre-requisites :
- Make sure that a class corresponding to the dataset exists in the project.
- Make sure that you have downloaded and stored the dataset inside a specific parent folder (we call it base_path)
NOTE: If 1. is not true, please look down at pre-requisites for developers.
- Give a reading to StarsDatasets.py.
- Read the inheriting class implementations for different datasets to get a clear idea of how the framework is written.
- Always assume that a user will provide a list of image filenames (with full file names) for which bounding box annotations are required.
- MAKE SURE that once you write an implementation of your own dataset, you update an example in test.py and send a pull request to update the project and make it more useful to a wider community.
Please see test.py